Prompt Library Manager

---
name: prompt-library-manager
description: Build and maintain a team's prompt library — copy-paste starters with bracketed variables, organised build-once vs run-daily, promoted to skills when they harden. Use when asked to organise prompts, build a prompt library or swipe file of prompts, stop the team rewriting the same prompt, standardise how everyone asks the AI, decide what becomes a skill vs stays a prompt, or someone says 'I know we solved this prompt before'. For hardening a prompt into a procedure, see skill-writer. For the always-loaded rules, see claude-md-writer.
metadata:
  version: 1.0.0
---

# Prompt library manager

Every handbook in the course ships one: a **prompt library of copy-paste
starters** — because a prompt that worked is an asset, and an asset nobody can
find gets rebuilt weekly at full price.

The organising insight, straight from the course's own libraries: **"Run 1 to 4
once to build the agent. Then live in 5 to 10."** Build prompts and daily
prompts are different species; file them apart.

## Before you start

| Input | Why |
| --- | --- |
| The prompts the team actually reuses today | The library starts from proven, not aspirational |
| Where the team already looks for things | A library nobody opens is a graveyard; put it where they are |
| The agent(s) the prompts drive | Prompts reference skills and files by name — names must match |

## Step 1 — File each prompt as build-once or run-daily

| Type | Nature | Course examples |
| --- | --- | --- |
| **Build** (run once) | Scaffold the agent, write CLAUDE.md, build the voice file, wire an API | Prompts 1–4 of the SEO library |
| **Run** (live here) | Validate a topic, publish a post, audit, plan a cluster, refresh | Prompts 5–10 |

A library that mixes them buries the six prompts someone needs daily under the
four they needed once.

## Step 2 — Write each entry in the house shape

```
### <NUMBER> · <WHAT IT DOES, AS A VERB PHRASE>
<the prompt, verbatim, with [BRACKETED VARIABLES] for everything that changes>
```

The course's conventions, extracted:

- **Bracketed variables** for the operator's specifics: `[BRAND]`, `[domain]`,
  `[topic]`, `[URL]`. The prompt ships ready to fill, never pre-filled with one
  person's details.
- **Constraints ride inside the prompt** — "8 to 15 cluster posts with validated
  volume", "do not apply anything until I approve it". The prompt carries its
  own guardrails so they survive the paste.
- **Numbered, short titles** — people cite "run prompt 6", not a paragraph.

## Step 3 — Promote what hardens

A prompt is a seed. When one is pasted weekly, has grown its own checklist, or
keeps needing the same corrections — **it has outgrown the library**:

| Signal | Destination |
| --- | --- |
| Same prompt + same follow-up corrections every time | A skill (skill-writer) — the corrections become steps |
| Prompt now 40 lines with rules inside | A skill — it already is one, unfiled |
| Stable decision the prompt keeps restating | CLAUDE.md — save decisions, not conversations |
| Used once, never again | Archive — the library is for reuse |

The library is the nursery; skills are where prompts graduate.

## Step 4 — Keep it alive

- One owner. Libraries without an owner fork into personal copies within a month.
- Every entry earns its place by use — prune on a schedule, and archive rather
  than delete.
- New entry rule: it was used successfully **twice** before it enters. Once is
  an experiment.

## Output

```
# Prompt library: <team/agent>

## Build (run once)          ## Run (live here)
1 · <title>                  5 · <title>
2 · <title>                  6 · <title>
...                          ...

Entries: <n> (each: numbered title + verbatim prompt + [VARIABLES])
Promoted this pass: <prompt → skill/CLAUDE.md, with reason>
Archived: <unused entries>
Owner: <name> · Next prune: <date>
```

## When it breaks

| What you see | What it means | The fix |
| --- | --- | --- |
| Team still rewrites known prompts | Library not where they work | Move it to the tool they already open |
| Everyone's copy has drifted | No owner | One owner; personal forks merge back or die |
| Prompts arrive pre-filled with someone's brand | Variables not bracketed | [BRAND] the specifics; ship templates, not residue |
| The good prompt needs three corrections every run | It hardened; nobody promoted it | Step 3: the corrections become a skill's steps |
| Library is 90 entries, six get used | No pruning, no entry bar | Twice-used rule in; schedule the prune |
| Daily prompts buried under setup prompts | Build and run mixed | Split the two sections; daily lives on top |
| Guardrails ignored after pasting | Constraints lived outside the prompt | Move the rules inside the prompt text itself |

Never let the library become the system of record for *procedures*. A procedure
pretending to be a prompt gets pasted with drift; the moment rules accumulate,
promote it to a skill where the five questions protect it.

## Rules

- **Build and run stay separate**, because daily users should never scroll past
  scaffolding to find their six living prompts.
- **Variables are bracketed, always**, because a pre-filled prompt is one
  person's residue and a bracketed one is everyone's template.
- **Constraints ride inside the prompt**, because guardrails stored next to the
  prompt do not survive the paste.
- **Twice-used before entry, pruned on schedule**, because a library's value is
  its signal-to-noise, not its size.
- **Promote what hardens**, because the library is a nursery — a 40-line prompt
  with rules is a skill wearing the wrong clothes.

## Related skills

- **skill-writer** — where hardened prompts graduate.
- **claude-md-writer** — where stable decisions land instead of being re-pasted.
- **token-budgeter** — combined-brief discipline shapes good prompt entries.
- **agent-folder-scaffold** — commands/ is where per-agent prompt shortcuts live.